Measurement comes before adjudication: a legal machine should hand over a thread, not a verdict.

The oldest dream in legal AI is a machine Hercules: a system that reads the whole law and returns the right answer. It has two fatal flaws. It inherits every serious objection to automated adjudication, since a decision is an exercise of public power. And it presupposes that a determinate answer exists to be found, which is exactly what legal theory disputes in hard cases.

An early experiment made the second flaw concrete. Identical materials given to seven simulated judges with different temperament parameters produced scores ranging from 2 to 8 and split between dismissing the claim and letting it proceed, while the same judge re-run stayed consistent. Disposition, not noise, drove the outcomes. A machine that returned one of those answers as the answer would be concealing the others.

The pivot is from building the judge to measuring the maze. Ariadne holds verified law fixed, varies only judicial posture, subtracts the instrument's noise, and reports how open the question is, with an interval, a noise floor, the postures driving the dispersion and a ledger of rationales. It decides nothing. Its output is a contestability artifact for human review.

That ordering is a position about legitimacy as much as method. A tool that measures can be wrong in ways that are checkable and cheap: its conjectures are stated and falsifiable, its panel is declared, and its failures show up in negative controls. A tool that adjudicates can be wrong in ways that bind people. The continental tradition prized legal certainty for centuries without being able to measure it. Measuring it is useful work. Replacing the person who must decide is not the job.